Unnecessary softmax in accuracy?

[DrChainsaw]

Just stumbled on this as I was poking around:

function accuracy(y_pred, y)
    return mean(onecold(cpu(softmax(y_pred))) .== onecold(cpu(y)))
end

That softmax should not be needed as onecold afaik only does argmax + some plumbing under the hood and softmax should not change the outcome of that.